computer-agent / e2bqwen.py
m-ric's picture
m-ric HF Staff
Re-add images in inference. Inference endpoint model still broken context length
13ad481
raw
history blame
22.5 kB
import os
import time
import base64
from io import BytesIO
from textwrap import dedent
from typing import Any, Dict, List, Optional, Tuple
import json
# HF API params
from huggingface_hub import InferenceClient
# E2B imports
from e2b_desktop import Sandbox
from PIL import Image
# SmolaAgents imports
from smolagents import CodeAgent, tool, HfApiModel
from smolagents.memory import ActionStep
from smolagents.models import ChatMessage, MessageRole, Model
from smolagents.monitoring import LogLevel
E2B_SYSTEM_PROMPT_TEMPLATE = """You are a desktop automation assistant that can control a remote desktop environment.
On top of performing computations in the Python code snippets that you create, you only have access to these tools to interact with the desktop, no additional ones:
{%- for tool in tools.values() %}
- {{ tool.name }}: {{ tool.description }}
Takes inputs: {{tool.inputs}}
Returns an output of type: {{tool.output_type}}
{%- endfor %}
The desktop has a resolution of <<resolution_x>>x<<resolution_y>>.
IMPORTANT:
- Remember the tools that you have as those can save you time, for example open_url to enter a website rather than searching for the browser in the OS.
- Whenever you click, MAKE SURE to click in the middle of the button, text, link or any other clickable element. Not under, not on the side. IN THE MIDDLE. In menus it is always better to click in the middle of the text rather than in the tiny icon. Calculate extremelly well the coordinates. A mistake here can make the full task fail.
- To navigate the desktop you should open menus and click. Menus usually expand with more options, the tiny triangle next to some text in a menu means that menu expands. For example in Office in the Applications menu expands showing presentation or writing applications.
- Always analyze the latest screenshot carefully before performing actions. If you clicked somewhere in the previous action and in the screenshot nothing happened, make sure the mouse is where it should be. Otherwise you can see that the coordinates were wrong.
You must proceed step by step:
1. Understand the task thoroughly
2. Break down the task into logical steps
3. For each step:
a. Analyze the current screenshot to identify UI elements
b. Plan the appropriate action with precise coordinates
c. Execute ONE action at a time using the proper tool
d. Wait for the action to complete before proceeding
After each action, you'll receive an updated screenshot. Review it carefully before your next action.
COMMAND FORMAT:
Always format your actions as Python code blocks. For example:
```python
click(250, 300)
```<end_code>
TASK EXAMPLE:
For a task like "Open a text editor and type 'Hello World'":
1- First, analyze the screenshot to find the Applications menu and click on it being very precise, clicking in the middle of the text 'Applications':
```python
click(50, 10)
```<end_code>
2- Remembering that menus are navigated through clicking, after analyzing the screenshot with the applications menu open we see that a notes application probably fits in the Accessories section (we see it is a section in the menu thanks to the tiny white triangle after the text accessories). We look for Accessories and click on it being very precise, clicking in the middle of the text 'Accessories'. DO NOT try to move through the menus with scroll, it won't work:
```python
click(76, 195)
```<end_code>
3- Remembering that menus are navigated through clicking, after analyzing the screenshot with the submenu Accessories open, look for 'Text Editor' and click on it being very precise, clicking in the middle of the text 'Text Editor':
```python
click(241, 441)
```<end_code>
4- Once Notepad is open, type the requested text:
```python
type_text("Hello World")
```<end_code>
5- Task is completed:
```python
final_answer("Done")
```<end_code>
Remember to:
Always wait for appropriate loading times
Use precise coordinates based on the current screenshot
Execute one action at a time
Verify the result before proceeding to the next step. If you repeated an action already without effect, it means that this action is useless: don't repeat it and try something else.
Use click to move through menus on the desktop and scroll for web and specific applications.
REMEMBER TO ALWAYS CLICK IN THE MIDDLE OF THE TEXT, NOT ON THE SIDE, NOT UNDER.
"""
class E2BVisionAgent(CodeAgent):
"""Agent for e2b desktop automation with Qwen2.5VL vision capabilities"""
def __init__(
self,
model: HfApiModel,
data_dir: str,
desktop: Sandbox,
tools: List[tool] = None,
max_steps: int = 200,
verbosity_level: LogLevel = 4,
planning_interval: int = 10,
log_file = None,
**kwargs
):
self.desktop = desktop
self.data_dir = data_dir
self.log_path = log_file
self.planning_interval = planning_interval
# Initialize Desktop
self.width, self.height = self.desktop.get_screen_size()
print(f"Screen size: {self.width}x{self.height}")
# Set up temp directory
os.makedirs(self.data_dir, exist_ok=True)
print(f"Screenshots and steps will be saved to: {self.data_dir}")
# Initialize base agent
super().__init__(
tools=tools or [],
model=model,
max_steps=max_steps,
verbosity_level=verbosity_level,
planning_interval = self.planning_interval,
**kwargs
)
self.prompt_templates["system_prompt"] = E2B_SYSTEM_PROMPT_TEMPLATE.replace("<<resolution_x>>", str(self.width)).replace("<<resolution_y>>", str(self.height))
# Add screen info to state
self.state["screen_width"] = self.width
self.state["screen_height"] = self.height
# Add default tools
self.logger.log("Setting up agent tools...")
self._setup_desktop_tools()
self.step_callbacks.append(self.take_screenshot_callback)
self.final_answer_checks = [self.store_metadata_to_file]
def _setup_desktop_tools(self):
"""Register all desktop tools"""
@tool
def click(x: int, y: int) -> str:
"""
Performs a left-click at the specified coordinates
Args:
x: The x coordinate (horizontal position)
y: The y coordinate (vertical position)
"""
self.desktop.move_mouse(x, y)
self.desktop.left_click()
self.logger.log(self.log_path, f"Clicked at coordinates ({x}, {y})")
return f"Clicked at coordinates ({x}, {y})"
@tool
def right_click(x: int, y: int) -> str:
"""
Performs a right-click at the specified coordinates
Args:
x: The x coordinate (horizontal position)
y: The y coordinate (vertical position)
"""
self.desktop.move_mouse(x, y)
self.desktop.right_click()
self.logger.log(self.log_path, f"Right-clicked at coordinates ({x}, {y})")
return f"Right-clicked at coordinates ({x}, {y})"
@tool
def double_click(x: int, y: int) -> str:
"""
Performs a double-click at the specified coordinates
Args:
x: The x coordinate (horizontal position)
y: The y coordinate (vertical position)
"""
self.desktop.move_mouse(x, y)
self.desktop.double_click()
self.logger.log(self.log_path, f"Double-clicked at coordinates ({x}, {y})")
return f"Double-clicked at coordinates ({x}, {y})"
@tool
def move_mouse(x: int, y: int) -> str:
"""
Moves the mouse cursor to the specified coordinates
Args:
x: The x coordinate (horizontal position)
y: The y coordinate (vertical position)
"""
self.desktop.move_mouse(x, y)
self.logger.log(self.log_path, f"Moved mouse to coordinates ({x}, {y})")
return f"Moved mouse to coordinates ({x}, {y})"
@tool
def type_text(text: str, delay_in_ms: int = 75) -> str:
"""
Types the specified text at the current cursor position.
Args:
text: The text to type
delay_in_ms: Delay between keystrokes in milliseconds
"""
self.desktop.write(text, delay_in_ms=delay_in_ms)
self.logger.log(self.log_path, f"Typed text: '{text}'")
return f"Typed text: '{text}'"
@tool
def press_key(key: str) -> str:
"""
Presses a keyboard key (e.g., "Return", "tab", "ctrl+c")
Args:
key: The key to press (e.g., "Return", "tab", "ctrl+c")
"""
if key == "enter":
key = "Return"
self.desktop.press(key)
self.logger.log(self.log_path, f"Pressed key: {key}")
return f"Pressed key: {key}"
@tool
def go_back() -> str:
"""
Goes back to the previous page in the browser.
Args:
"""
self.desktop.press(["alt", "left"])
self.logger.log(self.log_path, "Went back one page")
return "Went back one page"
@tool
def drag_and_drop(x1: int, y1: int, x2: int, y2: int) -> str:
"""
Clicks [x1, y1], drags mouse to [x2, y2], then release click.
Args:
x1: origin x coordinate
y1: origin y coordinate
x2: end x coordinate
y2: end y coordinate
"""
self.desktop.drag([x1, y1], [x2, y2])
message = f"Dragged and dropped from [{x1}, {y1}] to [{x2}, {y2}]"
self.logger.log(self.log_path, message)
return message
@tool
def scroll(direction: str = "down", amount: int = 1) -> str:
"""
Uses scroll button: this could scroll the page or zoom, depending on the app. DO NOT use scroll to move through linux desktop menus.
Args:
direction: The direction to scroll ("up" or "down"), defaults to "down"
amount: The amount to scroll. A good amount is 1 or 2.
"""
self.desktop.scroll(direction=direction, amount=amount)
self.logger.log(self.log_path, f"Scrolled {direction} by {amount}")
return f"Scrolled {direction} by {amount}"
@tool
def wait(seconds: float) -> str:
"""
Waits for the specified number of seconds. Very useful in case the prior order is still executing (for example starting very heavy applications like browsers or office apps)
Args:
seconds: Number of seconds to wait, generally 3 is enough.
"""
time.sleep(seconds)
self.logger.log(self.log_path, f"Waited for {seconds} seconds")
return f"Waited for {seconds} seconds"
@tool
def open_url(url: str) -> str:
"""
Directly opens a browser with the specified url, saves time compared to clicking in a browser and going through the initial setup wizard.
Args:
url: The URL to open
"""
# Make sure URL has http/https prefix
if not url.startswith(("http://", "https://")):
url = "https://" + url
self.desktop.open(url)
# Give it time to load
time.sleep(2)
self.logger.log(self.log_path, f"Opening URL: {url}")
return f"Opened URL: {url}"
# Register the tools
self.tools["click"] = click
self.tools["right_click"] = right_click
self.tools["double_click"] = double_click
self.tools["move_mouse"] = move_mouse
self.tools["type_text"] = type_text
self.tools["press_key"] = press_key
self.tools["scroll"] = scroll
self.tools["wait"] = wait
self.tools["open_url"] = open_url
self.tools["go_back"] = go_back
self.tools["drag_and_drop"] = drag_and_drop
def store_metadata_to_file(self, final_answer, memory) -> None:
metadata_path = os.path.join(self.data_dir, "metadata.json")
output = {}
# THIS ERASES IMAGES FROM MEMORY, USE WITH CAUTION
for memory_step in self.memory.steps:
if getattr(memory_step, "observations_images", None):
memory_step.observations_images = None
a = open(metadata_path,"w")
a.write(json.dumps(self.write_memory_to_messages()))
a.close()
return True
def take_screenshot_callback(self, memory_step: ActionStep, agent=None) -> None:
"""Callback that takes a screenshot + memory snapshot after a step completes"""
self.logger.log(self.log_path, "Analyzing screen content...")
current_step = memory_step.step_number
time.sleep(2.0) # Let things happen on the desktop
screenshot_bytes = self.desktop.screenshot()
image = Image.open(BytesIO(screenshot_bytes))
# Create a filename with step number
screenshot_path = os.path.join(self.data_dir, f"step_{current_step:03d}.png")
image.save(screenshot_path)
print(f"Saved screenshot for step {current_step} to {screenshot_path}")
for (
previous_memory_step
) in agent.memory.steps: # Remove previous screenshots from logs for lean processing
if (
isinstance(previous_memory_step, ActionStep)
and previous_memory_step.step_number <= current_step - 2
):
previous_memory_step.observations_images = None
# Add to the current memory step
memory_step.observations_images = [image.copy()] # This takes the original image directly.
# memory_step.observations_images = [screenshot_path] # IF YOU USE THIS INSTEAD OF ABOVE, LAUNCHING A SECOND TASK BREAKS
def close(self):
"""Clean up resources"""
if self.desktop:
print("Stopping e2b stream and killing sandbox...")
self.desktop.stream.stop()
self.desktop.kill()
print("E2B sandbox terminated")
class QwenVLAPIModel(Model):
"""Model wrapper for Qwen2.5VL API with fallback mechanism"""
def __init__(
self,
hf_base_url,
model_path: str = "Qwen/Qwen2.5-VL-72B-Instruct",
provider: str = "hyperbolic",
hf_token: str = None,
):
super().__init__()
self.model_id = model_path
self.hf_base_url = hf_base_url
self.dedicated_endpoint_model = HfApiModel(
hf_base_url,
token=hf_token
)
self.fallback_model = HfApiModel(
model_path,
provider=provider,
token=hf_token,
)
def __call__(
self,
messages: List[Dict[str, Any]],
stop_sequences: Optional[List[str]] = None,
**kwargs
) -> ChatMessage:
try:
return self.dedicated_endpoint_model(messages, stop_sequences, **kwargs)
except Exception as e:
print(f"HF endpoint failed with error: {e}. Falling back to hyperbolic.")
# Continue to fallback
try:
return self.fallback_model(messages, stop_sequences, **kwargs)
except Exception as e:
raise Exception(f"Both endpoints failed. Last error: {e}")
# class QwenVLAPIModel(Model):
# """Model wrapper for Qwen2.5VL API with fallback mechanism"""
# def __init__(
# self,
# model_path: str = "Qwen/Qwen2.5-VL-72B-Instruct",
# provider: str = "hyperbolic",
# hf_token: str = None,
# hf_base_url: str = "https://n5wr7lfx6wp94tvl.us-east-1.aws.endpoints.huggingface.cloud"
# ):
# super().__init__()
# self.model_path = model_path
# self.model_id = model_path
# self.provider = provider
# self.hf_token = hf_token
# self.hf_base_url = hf_base_url
# # Initialize hyperbolic client
# self.hyperbolic_client = InferenceClient(
# provider=self.provider,
# )
# assert not self.hf_base_url.endswith("/v1/"), "Enter your base url without '/v1/' suffix."
# # Initialize HF OpenAI-compatible client if token is provided
# self.hf_client = None
# from openai import OpenAI
# self.hf_client = OpenAI(
# base_url=self.hf_base_url + "/v1/",
# api_key=self.hf_token
# )
# def __call__(
# self,
# messages: List[Dict[str, Any]],
# stop_sequences: Optional[List[str]] = None,
# **kwargs
# ) -> ChatMessage:
# """Convert a list of messages to an API request with fallback mechanism"""
# # Format messages once for both APIs
# formatted_messages = self._format_messages(messages)
# # First try the HF endpoint if available - THIS ALWAYS FAILS SO SKIPPING
# try:
# completion = self._call_hf_endpoint(
# formatted_messages,
# stop_sequences,
# **kwargs
# )
# print("SUCCESSFUL call of inference endpoint")
# return ChatMessage(role=MessageRole.ASSISTANT, content=completion)
# except Exception as e:
# print(f"HF endpoint failed with error: {e}. Falling back to hyperbolic.")
# # Continue to fallback
# # Fallback to hyperbolic
# try:
# return self._call_hyperbolic(formatted_messages, stop_sequences, **kwargs)
# except Exception as e:
# raise Exception(f"Both endpoints failed. Last error: {e}")
# def _format_messages(self, messages: List[Dict[str, Any]]):
# """Format messages for API requests - works for both endpoints"""
# formatted_messages = []
# for msg in messages:
# role = msg["role"]
# content = []
# if isinstance(msg["content"], list):
# for item in msg["content"]:
# if item["type"] == "text":
# content.append({"type": "text", "text": item["text"]})
# elif item["type"] == "image":
# # Handle image path or direct image object
# if isinstance(item["image"], str):
# # Image is a path
# with open(item["image"], "rb") as image_file:
# base64_image = base64.b64encode(image_file.read()).decode("utf-8")
# else:
# # Image is a PIL image or similar object
# img_byte_arr = BytesIO()
# base64_image = base64.b64encode(img_byte_arr.getvalue()).decode("utf-8")
# content.append({
# "type": "image_url",
# "image_url": {
# "url": f"data:image/png;base64,{base64_image}"
# }
# })
# else:
# # Plain text message
# content = [{"type": "text", "text": msg["content"]}]
# formatted_messages.append({"role": role, "content": content})
# return formatted_messages
# def _call_hf_endpoint(self, formatted_messages, stop_sequences=None, **kwargs):
# """Call the Hugging Face OpenAI-compatible endpoint"""
# # Extract parameters with defaults
# max_tokens = kwargs.get("max_new_tokens", 4096)
# temperature = kwargs.get("temperature", 0.7)
# top_p = kwargs.get("top_p", 0.9)
# stream = kwargs.get("stream", False)
# completion = self.hf_client.chat.completions.create(
# model="tgi", # Model name for the endpoint
# messages=formatted_messages,
# max_tokens=max_tokens,
# temperature=temperature,
# top_p=top_p,
# stream=stream,
# stop=stop_sequences
# )
# if stream:
# # For streaming responses, return a generator
# def stream_generator():
# for chunk in completion:
# yield chunk.choices[0].delta.content or ""
# return stream_generator()
# else:
# # For non-streaming, return the full text
# return completion.choices[0].message.content
# def _call_hyperbolic(self, formatted_messages, stop_sequences=None, **kwargs):
# """Call the hyperbolic API"""
# completion = self.hyperbolic_client.chat.completions.create(
# model=self.model_path,
# messages=formatted_messages,
# max_tokens=kwargs.get("max_new_tokens", 4096),
# temperature=kwargs.get("temperature", 0.7),
# top_p=kwargs.get("top_p", 0.9),
# stop=stop_sequences
# )
# # Extract the response text
# output_text = completion.choices[0].message.content
# return ChatMessage(role=MessageRole.ASSISTANT, content=output_text)
# def to_dict(self) -> Dict[str, Any]:
# """Convert the model to a dictionary"""
# return {
# "class": self.__class__.__name__,
# "model_path": self.model_path,
# "provider": self.provider,
# "hf_base_url": self.hf_base_url,
# # We don't save the API keys for security reasons
# }
# @classmethod
# def from_dict(cls, data: Dict[str, Any]) -> "QwenVLAPIModel":
# """Create a model from a dictionary"""
# return cls(
# model_path=data.get("model_path", "Qwen/Qwen2.5-VL-72B-Instruct"),
# provider=data.get("provider", "hyperbolic"),
# hf_base_url=data.get("hf_base_url", "https://s41ydkv0iyjeokyj.us-east-1.aws.endpoints.huggingface.cloud"),
# )